# Automatic memory review Open top-level chats, discussions and normal/Enno runs review newly completed human turns without waiting for the session to close. The default interval is 8 turns. A review may add a project candidate, update a candidate owned by this feature, leave it unchanged, or hold it for later evaluation. An empty result is a successful review. Saving a memory does not make it verified. The next human request selects relevant memories using its current question and the existing scope, capability and memory-use gates. Active Enno work keeps its existing refresh owner and budgets. Saving and retrieval are separate steps; not every saved memory belongs in every response. ## Configuration and cost These are plugin configuration fields (under the existing `kiokuko-dsh` entry): ```yaml memoryReview: mode: active turnInterval: 8 minimumTurns: 4 dailyCalls: 12 maxInputBytes: 32768 maxOutputTokens: 2048 timeoutMs: 30000 boundaryFlush: true notifications: changes ``` `active` can write candidates. `observe` makes the same paid model call and records decisions but does not change entries. `off` stops periodic/manual review; it does not disable ordinary/Deep Finalizer or Evolution. Those have separate responsibilities and budgets. `minimumTurns` must not exceed `turnInterval`. A nonterminal session boundary can flush at least four completed turns. Idle alone does not lower the eight-turn threshold. A job retains the provider/model from its confirmed native request. Unknown context capacity, missing native indexed reading/flush, or unavailable model service prevents dispatch. There is no fallback provider or separate small model. One dispatched job calls the DSH stream API once with no tools; a provider's internal HTTP retries are outside that worker-level guarantee. Failures after dispatch consume the UTC-day allowance and are not automatically resent. Manual retry creates a new budgeted child job. Settings have a persistent generation shared by hosts. An older host cannot silently overwrite a newer generation. A cold start with `mode: off` can disable saved active settings. If requested and effective settings differ, explicitly set the desired mode in the current project: ```text /kioku-memory-review mode off /kioku-memory-review mode observe /kioku-memory-review mode active /kioku-memory-review status --json ``` Other fields apply on first admission or a live configuration update. On cold restart, persisted settings remain authoritative; status exposes requested and effective modes. Mode changes do not retry cancelled or uncertain jobs. ## Inspect and recover Status, mode and exclusion commands also work before the first task is admitted. Run manual review and retry commands inside an admitted top-level session: ```text /kioku-memory-review status /kioku-memory-review status --json /kioku-memory-review run /kioku-memory-review retry /kioku-memory-review retry-finalizer ``` `run` reviews unscheduled, completed turns. It does not resend an existing job. `retry` accepts unresolved held/rejected/cancelled work, including a completed job with held operations. Repeating it returns the same child job. Use that child's ID for another deliberate attempt. Retry preserves the original model, checks current evidence and candidates, and refuses to overwrite later adopted evidence with an older range. Missing source logs, saved exclusions, review off, and an active Finalizer prevent retry. Failed Finalizers retain their own three-attempt limit; `retry-finalizer` does not consume a review call. Status includes the scanned/scheduled/reviewed cursors, jobs and recovery availability, provider/model, reasons, operation counts, request bytes, duration, adoption wait, UTC-day calls and remaining allowance. Reported token usage stays `null` when the provider did not report it. Bytes and the conservative context admission bound are not measured tokens. Finalizer usage and Evolution calls are shown separately. Change notifications use a native session notice, not a new human turn. Status does not expose evidence text or raw exception messages. A terminal run hands remaining work to an actual compatible Finalizer reservation. Cancelled runs and failed runs without a consumer keep unresolved state; they are not reported as reviewed. Review and ordinary Finalizer share a database lease. Expired undispatched jobs can recover, while expired dispatched reviews remain held because the provider result is unknown. ## Exclude this conversation Before sensitive input, use: ```text /kioku-memory-review exclude session ``` The exact standalone phrases `この会話を保存しない` and `この会話を覚えないで` also exclude the entire session. Limited detection of `保存しない`, `覚えないで`, `do not remember` and `don't remember` holds automatic capture; quotes and hypothetical examples can therefore also cause a hold. Inputs over 1 MiB are held when full inspection is impractical. This is not complete recognition of arbitrary natural-language refusal. The saved exclusion applies to review, old and new ordinary Finalizers, Deep memory extraction, and Evolution jobs containing that session's evidence. It fences new dispatch and adoption and requests cancellation of local in-flight calls. It cannot retract data already sent. It does not stop normal chat/Deep execution, erase native logs or existing memories, or disable explicit memory tools. Exclusion survives restart; an excluded parent is inherited by a branch. Unknown lineage is held. This version has no partial exclusion or same-session unexclude command; start a new conversation for capture to resume. ## Evidence and retention Only explicit human sources and paired eligible native tool observations are new evidence. Assistant assertions, memory/control tool output and injected memories do not become independent evidence. Model instructions require preserving negation, conditions, corrections and temporary preferences; host checks enforce IDs, revisions, scope, ownership, secret rejection and trust. The model's semantic interpretation can still be wrong. Inspect important candidates before treating them as facts. Review input uses bounded immutable evidence manifests and candidate snapshots; it does not copy entire native history into Core. Large inputs split only at complete turn boundaries, carrying up to 8 KiB of earlier evidence as context that cannot independently support another memory. If required context exceeds that bound, the affected chunk remains held. Oversized individual units or secret-bearing units remain held. No event-body truncation is presented as complete evidence. Human-edited, verified, superseded and externally managed entries cannot be automatically updated. Deleted/superseded source effects remain tombstones. Finalizer v3 reconciles reviews and retains fact, decision, preference, lesson and reference support. Old jobs retain their v1/v2 contract. Episode extraction remains one per run and does not count periodic reviews as additional votes. Source logs use existing retention; eviction makes unresolved work `source_unavailable` rather than fetching or sending missing history silently. ## Verification Local verification on 2026-09-19: the full suite with the pinned DSH runtime finished with 828 passed, 0 failed and 28 environment-gated skips (856 tests). The new feature suite contains 29 tests, including native open-chat delivery, refusal notices, legacy v1/v2 exclusion and shared Finalizer lease recovery. Typecheck, build, publint and package closure/import checks passed. ```bash npm run test:memory-review KIOKUKO_DSH_PACKAGE_ROOT="$PWD/tests/fixtures/dsh-runtime/node_modules" npm run test:memory-review npm run test:benchmark:memory-review ``` The native integration tests exercise DSH Session, agent loop and registered provider adapter with a deterministic mock model. They verify saving in an open chat and retrieval on the next request with `ennoMemory` off/observe/active. Separate tests cover two OS processes sharing SQLite, budget rollover, restart, late results, exclusion, candidate protection and Finalizer reconciliation. The incremental benchmark uses 10,000 / 100,000 / 1,000,000 historical events and 100 new turns each. On 2026-09-19, each read 501 native indexed events and 2,000 range rows (classification, reservation, dispatch validation and adoption validation), buffering at most five new native events. Measured delta work was 339 / 280 / 220 ms in that local fixture, excluding seeding. This is not a claim of zero user-response latency or real-provider performance. Real-model extraction quality and usefulness in subsequent work are **not measured**: no explicit live evaluation connection was supplied. Structural fixture results do not establish semantic accuracy. Publication, installation and verification in the user's currently running DSH profile are separate.